wjaskowski / mastering-2048

An efficient reinforcement learning algorithm for learning a strategy for game 2048

About

The source code for running the experiments for the paper:

Wojciech Jaśkowski "Mastering 2048 with Delayed Temporal Coherence Learning, Multi-State Weight Promotion, Redundant Encoding and Carousel Shaping", IEEE Transactions on Computational Intelligence and AI in Games (accepted) [arXiv]

The code in this repo has been used to train the best 2048 controller, which (along with an efficient C++ implementation) is available in a separate repo.

Authors

Wojciech Jaśkowski

(Some code in this repository is due to Marcin Szubert and Paweł Liskowski)

Prerequisites

Java 8, Maven

Building

> mvn install:install-file -Dfile=lib/stilts.jar -DgroupId=uk.ac.starlink -DartifactId=stilts -Dversion=2.4 -Dpackaging=jar
> mvn package -Dmaven.test.skip=true

Running

Example:

> java -Xmx50g -Dlog4j.configuration=file:configs/tciaig-2048/log4j.properties -Dframework.properties=configs/tciaig-2048/42-33_tcl-0.5-0.5.properties -Dseed=123 -Dresults_dir=results/tcl/123 -jar cevo.jar

Note: some of the experiments require a lot of memory (32GB might not be enough).

Available config files

Disclaimer

This repo contains a lot of code irrelevant to 2048